Running a Local AI Stack on RTX 3090 Cost Just €2 in Electricity Over 30 Days

A developer running a local AI stack — including a permanent Whisper transcription service and an embedding model on an NVIDIA RTX 3090 — measured actual electricity costs over 30 days and found the GPU drew only €2.00 worth of power. The low cost is explained by the nature of inference workloads: the GPU averaged just 25 watts, sitting mostly idle at around 35 watts and spiking to roughly 120 watts only briefly during active transcription. The Whisper ASR service accounted for €1.76 of the total, while the embedding model cost just 23 cents for the entire month. The author notes that the dominant cost of local LLMs is hardware amortisation — an RTX 3090 purchased used for €800–1,200 works out to roughly €25 per month over three years — not electricity. The key takeaway is that a home inference stack behaves like a mostly idle appliance rather than a continuously loaded training rig.